基于生物视觉感知机理的高效图像处理技术研究

Information processing mechanisms within the biological vision system may provide strong inspiration to advance the theories of technologies in the field of the computer vision and computational modeling of biological vision system is quite helpful of better understanding of the working mechanisms of our brain, and also very helpful of developing brain-like information processing methods. In this project, based on our plentiful achievements obtained during the past years, we aim to develop efficient models inspired the visual mechanisms, especially at the level of retina, and also some important mechanisms involved in the mid- and high-level areas. We mainly care about the mechanisms underlying the various structures and functions of the non-classical receptive fields and the mechanisms underlying the dynamic adaptation involved in multiple levels. These mechanisms include the light and dark adaptation, contrast adaptation and chromatic adaptation involved in the retina; and the feedback control, selective attention, color constancy involved in the levels of V1, V2, V4, etc. Besides the developing of visual models, we will also conduct some essential physiological and psychophysical studies in order to uncover some unclear mechanisms that are important for developing efficient models. The ultra goal of the project within 3 years is to develop biological vision inspired models that can quite efficiently and effectively deal with the image processing problems like image denosing, enhancement, dehazing, luminance and color recovering, high-dynamic image rendering, color constancy, etc. This work is expected to provide novel ideas and models to potentially make breakthrough in theories and technologies of computer vision system.

生物视觉机理是计算机视觉理论与技术革新的重要源泉。生物视觉建模对于阐明大脑工作机理及发展脑智技术均意义重大。本项目以我们长期积累的相关研究成果为坚实基础，同时广泛吸收国际上在视觉机理及模型（特别是在视网膜的复杂结构及独特功能）方面的最新成果，将认知科学成果、机器感知方法与信息处理技术结合，以发展高效的图像/视频预处理技术为主要目标，以非经典感受野的多变结构和功能及视觉系统的自适应特性为核心，重点围绕视网膜层次的生物机理（如明暗适应、对比度适应、颜色适应等）进行数学建模；同时关注中、高级层次的相关视觉机制（如反馈、选择性注意）；并配以必要的实验研究。力求在若干关键理论和计算模型方面取得突破性进展，最终建立针对复杂场景（如强噪声、低对比度、大雾、颜色失真、高动态亮度范围等）的高效图像预处理技术（如去噪、去雾、增强、颜色矫正、亮度恢复等），为相关的计算机视觉应用提供坚实的技术支持。

本项目共发表或录用论文19篇，其中SCI论文15篇（其中包括IEEE T-PAMI在内的IEEE Transactions论文8篇），EI论文4篇（含国际顶级会议CVPR论文1篇）；授权中国发明专利6项，新申请中国发明专利9项。在生物视觉机理方面，通过电生理实验，研究了猫V1区两类代表性的神经元（兴奋性锥体细胞和抑制性中间神经元）对不同对比度图像信息在空间整合上的差异，揭示了自适应视觉信息加工的在神经元（感受野）层次的可能生物学基础。在基于生物视觉机理的高效图像处理技术方面：（1）通过模拟视网膜层次的亮度及颜色通路的交互作用机理，实现了场景去雾功能。（2）系统性地提出了图像颜色恒常的生物视觉机理和方法：我们模拟视网膜的多层次信息加工过程，建立了颜色恒常的视网膜神经网络模型，实现了对色偏场景的初步颜色矫正；通过引入视网膜光感受器层次的LMS颜色空间转换，建立了更加符合生理实际的视网膜-LGN-V1颜色恒常模型；基于眼动实验结果，提出了一种利用场景灰度信息实现多光源颜色恒常的方法。（3）完善了复杂场景物体轮廓检测的生物视觉机理和方法：分析了V1区中外周朝向选择性抑制神经元和外周非朝向选择性抑制神经元在轮廓检测中的不同作用；通过引入稀疏编码思想，对V1区朝向选择性双拮抗神经元的物体轮廓响应进行自适应调制，得到了更为合理的物体轮廓结果。（4）建立了生物视觉选择性注意计算模型：借鉴生物视觉选择注意中的引导搜索理论，建立了基于Bayes理论融合Bottom-up和Top-down信息的选择性注意计算框架，高效地实现了对复杂场景中显著目标的检测；进一步地，对驾驶环境中驾驶员与非驾驶员的眼动特征进行了详细研究，在基于Bottom-up的显著图计算模型基础上，融入了与交通环境高级认知特性相关的Top-down机制，实现了对驾驶员眼动搜索显著图的准确预测。